Various procedures for detecting and characterizing sources were tested by simulated X-ray images. We have concentrated our attention mainly on images with XMM-Newton specific characteristics, because the problems arising from its high sensitivity and relatively large PSF are new and challenging.

We have analyzed the detection rate and the recovery of all characteristics of the input objects: flux, positional accuracy, extent measurements and the recovery of the input $\log N-\log S$ relation. We have also investigated confusion problems in large exposures.

Concerning detection rate and characteristics reconstruction, we have shown that the VTPDETECT implementation of the Voronoi Tessellation and Percolation method is not suited to XMM-Newton images. EWAVELET provides very good detection rate and photometric reconstruction for point-like sources after a simple correction, but shows unreliable results for extended sources.

One of the best methods for point-like source detection and flux measurements is EMLDETECT but we stress again that the PSF model used for the ML procedure needs to be close to the image PSF for the most accurate photometry. Serious drawbacks are the relatively large number of spurious detections as well as the splitting of the extended sources, which we were not able to suppress even with 6 simultaneous PSF profile fits in the multi-PSF mode; this seriously hampers the analysis of the extended sources.

WAVDETECT is a flexible method giving good detections even in some complicated cases. But, here again, spurious detections are quite numerous. WAVDETECT does not assume a PSF model but requires the PSF size as a function of the encircled energy fraction and the off-axis distance in order to define the object detection cell. However, the way the detection cell is defined leads to bad photometry for extended objects.

Our choice is the MR/1+SE method. The mixed approach involving first a multiresolution iterative threshold filtering of the raw image followed by detection and analysis with SExtractor. Our tests have shown that this is the best strategy for detecting and characterizing both point-like and extended objects. Even though this mixed approach consists of two distinct steps, it is one of the fastest procedures (Table 9), allowing easy checks of different stages in the analysis (filtering, detection, photometry).

Table 9: CPU times for performing Test 5 for 100 ks. The procedures were run on a Pentium III-Xeon, 550 MHz, Linux. G+SE and MR/1+SE CPU time is the total of filtering and detection passes. For EMLDETECT the number of input objects is given
Procedure Number of CPU time

detections [min]

EMLDETECT 528 12.0

EWAVELET 364 0.4

MR/1+SE 370 1.9

G+SE 365 0.1

WAVDETECT 378 10.3

VTPDETECT 1307 10.7

**Table 9:** CPU times for performing Test 5 for 100 ks. The procedures were run on a Pentium III-Xeon, 550 MHz, Linux. `G+SE` and `MR/1+SE` CPU time is the total of filtering and detection passes. For `EMLDETECT` the number of input objects is given
Procedure	Number of	CPU time
	detections	[min]
`EMLDETECT`	528	12.0
`EWAVELET`	364	0.4
`MR/1+SE`	370	1.9
`G+SE`	365	0.1
`WAVDETECT`	378	10.3
`VTPDETECT`	1307	10.7

Without blending or confusion effects, the photometry is accurate within 10-20% for both point-like and extended objects. This uncertainty can be regarded as an intrinsic error due to the Poissonian nature of the X-ray images. For extended objects, only the MR/1+SE method gives satisfactory photometric results.

Blending between extended and point-like sources is quite serious at separations below $\sim$ $30\hbox {$^{\prime \prime }$ }$ . Better results for photometry may eventually be obtained if the intrinsic shape of the extended objects is known, and if the two objects are detected. However, in most of the cases with small separation, there is no indication of blending - which is a dangerous situation for flux reconstruction. In such cases, there may exist some spectral signatures of the effect.

The identification process of X-ray sources relies on their positional accuracy. We have shown that for objects with more than 100 counts in 10 ks exposure images and within the inner $10\hbox {$^\prime $ }$ of the field-of-view, the one sigma positional error is of the order of one half of the FWHM of the PSF ( $\approx$ $3{-}4\hbox{$^{\prime\prime}$ }$ , Table 6). For extended objects, because of their shallower profiles and depending on the number of photons and the off-axis distance, the detected centre could even be at about $15\hbox {$^{\prime \prime }$ }$ from its input position.

Comparing series of simulations with 100 ks and 10 ks in two energy bands - [0.5-2] and [2-10] keV, we show that the effects of confusion and completeness are absent for 10 ks, but quite significant for 100 ks in the lower energy band. Moreover, for faint fluxes, these effects tend to be masked by the large number of spurious detections with EMLDETECT. Although this method seems to give correct results for the $\log N-\log S$ down to fainter fluxes than MR/1+SE, in real situations it is impossible to asses the contribution of the numerous spurious detections. From our simulations, we estimate that about 60-65% of the sources are lost between 3 10^-16 and 6 10^-16 erg/s/cm² for a 100 ks exposure with the current best method (MR/1+SE).

One of the most important conclusions that will have deep cosmological impact concerns the detection and classification of extended objects. We have shown that the MR/1+SE mixed approach is capable of detecting galaxy cluster-like objects with moderate luminosity ( $L_{[2{-}10]\,{\rm keV}} \sim 3~10^{44}$ erg/s) at redshifts 1.5 < z< 2 in 10 ks XMM-Newton simulated images. A criteria based on the half-light radius and the stellarity index classifies them correctly, with a confidence level greater than 98%.

9 Conclusions